Methods & data

The steps

1. Choose a variable and a region. Each NASA dataset is stored on a 0.5° grid (about 55 km). The region's monthly value is the average over its grid cells, weighted by cell area; months with less than half the region observed are skipped.
2. Build a time series. Months are combined into yearly (or seasonal) totals or averages. A year counts only if at least 10 of its 12 months have data; a few missing months are filled with that month's usual value. For monthly analyses the seasonal cycle is removed.
3. Measure the trend. Sen's slope gives how fast the value changes per decade, and a least-squares line gives a 95% range of plausible trends.
4. Test significance. The Mann–Kendall test asks whether the rise or fall is more consistent than chance would give. Both tests are corrected for autocorrelation. MegavirTrace calls a trend significant only when the Mann–Kendall test (p < 0.05) and the 95% range of the linear trend (which must exclude zero) agree; when only one does, it reports mixed evidence.
5. Map where. The cell map repeats the test in every grid cell and uses a false-discovery-rate correction so the highlighted cells are not mostly chance.

In simulations of 25-year records with no real trend, this rule calls about 3% of them significant when years are independent and about 7% when each year strongly resembles the last (lag-1 autocorrelation 0.5). A real trend of one standard deviation per decade is detected about 84% of the time with independent years.

Extra clues for detectives

• Unusual years are more than twice the usual scatter away from the trend line (open rings on the chart).
• Was this month normal? compares one month with the same month in every other year: its difference from the usual value, divided by how much that month normally varies.
• Plot twists: if the data rose and then fell (or the reverse), a single straight line hides it. We search for the year where two opposite trends fit much better than one. Because that year is searched for, treat it as a lead, not a proof.
• Second opinion: for air temperature we compare two independent NASA sources, a reanalysis model (MERRA-2) and thermometers (GISTEMP), for the same place and years.
• Opposite trends finder tests every preset region. Testing many regions at once lets a few look significant by chance, so pairs are marked robust only if they survive a false-discovery-rate correction.
• Raw vs corrected maps: switch the cell map between testing each cell on its own and correcting for testing hundreds of cells at once, and see how many “significant” cells were just chance.
• Every result has a fingerprint: the same analysis on the same data always gives the same one. See Check our work for comparisons with official numbers.
• Fact cards quote published studies about a place, word for word, with the source. A fact is shown only after a team member has checked the quote in the source (review page).

What a trend can and cannot tell you

• A trend shows that something changed, not why. Causes need more evidence, such as physical understanding or other variables.
• "Not significant" does not mean "no change": often the record is simply too short. Try a longer period.
• Only straight-line (linear) trends are tested. Real changes can speed up, slow down or reverse.
• Satellite and model products have their own uncertainties, and a dataset version can change over time.

Variables

NASA data

Glossary

Trend

The overall direction a measurement moves in over many years, ignoring the ups and downs from one year to the next.

Sen's slope

The middle value of the slopes between every pair of years. One extreme year cannot pull it far, so it is a robust estimate of how fast something changes.

Linear trend (least squares)

The straight line that fits the data best. Its 95% range shows how steep the line could plausibly be.

Mann–Kendall test

Counts how often later years are higher than earlier years. If that happens far more (or less) often than chance would give, the trend is called significant. It does not assume the data follow a bell curve.

p-value

How surprising the data would be if there were no trend at all. Below 0.05 (5%) we call the trend statistically significant.

Statistically significant

Unlikely to be explained by random year-to-year variation alone. It does not say how important the change is, or what caused it.

Autocorrelation

When one year resembles the next (a wet year tends to follow a wet year). It makes trends look more certain than they are, so the tests are corrected for it.

Anomaly

The difference from the usual value for that time of year. Using anomalies removes the seasonal cycle so the trend can be seen.

False discovery rate

When we test hundreds of map cells at once, some will look significant by chance. The FDR correction keeps the expected share of false alarms among the highlighted cells at 10%.